Dear Doc:
I read that putting information into an artificial intelligence (AI) chatbot such as ChatGPT can break attorney-client privilege in some instances. Are there other things that we should worry about when we use AI?
Signed,
Concerned Person Who is Definitely Not a Robot
Dear CP:
The case you read about is United States v. Heppner. In that case, Judge Rakoff of the United States District Court for the Southern District of New York ruled against attorney-client privilege in information input to a public AI chatbot. The client asked Claude for answers about a legal matter. Judge Rakoff found that “Claude is not an attorney.” (Duh, says the Doc!) and thus whatever Claude was told was not protected by attorney-client privilege. But the Judge went even further, finding that the information was neither “intended to be,” nor was it “in fact,” “kept confidential.” He also determined that “Heppner could have had no ‘reasonable expectation of confidentiality in his communications with Claude.” Finally, he ruled that the use of Claude was not “for the purpose of obtaining legal advice.”
The bottom line here is that while your attorney may, in some instances, use a specialized legal AI system in her work, you, as her client, had better not put your confidential information into a chatbot.
The Doc takes this even further… every confidentiality and nondisclosure agreement (NDA) must now be revised because AI is becoming ubiquitous in business. Some organizations put every word that is written, typed or spoken into an AI system. If your confidential information, shared under an NDA, ends up in such a system and may thus used to train the AI itself, any hope of it remaining secret may instantly vanish. Thus, your NDA should specifically prohibit loading such confidential information into both public and corporate AI repositories. Complying with such a prohibition may prove exceptionally difficult for certain organizations, so counsel should be on the lookout for inadvertent disclosures.
Tools including ChatGPT, Microsoft Copilot, Claude and Google Gemini have become as routine as email in many professional organizations. As one important example, consider the common practice in the field of mergers and acquisitions of having a “data room”, which is a shared repository of documents used in the due diligence evaluation of the proposed transaction. In these deals, associates use large language models to summarize due diligence materials. Analysts construct financial projections using AI-powered platforms to test their underlying assumptions. Real estate brokers use chatbots to draft lease abstracts or compare sales data.
Most NDAs in use today were drafted before AI appeared. Their confidentiality obligations typically restrict disclosure to “third parties” or limit use to a defined “purpose.” The agreements were written with human recipients in mind—employees, advisors, affiliates—and sometimes require that those recipients be bound by obligations of confidentiality. The agreements do not address the involvement of machines. When a professional inputs deal-sensitive information into an AI tool, that information leaves the organization’s control and enters a third-party system. This may constitute a breach of the agreement. The information is shared outside the circle of permitted recipients, transmitted to infrastructure owned and operated by a technology company, and becomes subject to that company’s terms of service, which the disclosing party never agreed to.
What is needed is a new generation of NDA provisions designed specifically to address AI-related disclosure risks. These provisions typically include one or more of the following:
- Express prohibitions on inputting confidential information into any AI tool, whether or not the tool trains on user inputs.
- New definitions of disclosure that expressly encompass submission of information to automated systems, machine learning models, or AI-powered platforms.
- Provisions for enterprise AI tools that maintain strict data isolation, disable training on user inputs, and comply with strict security standards.
- Representations regarding internal AI policies, requiring the receiving party to confirm that it maintains organizational controls governing employee use of AI tools in connection with confidential information.
If your standard NDA does not contain provisions governing these AI-related issues, it’s time to revise them. Give the attorneys at LW&H a call to get started.
Until next month,
The “Doc”
— Lawrence A. Husick, Esq.